weekend news： 1.Tencent just admitted its AI agents went rogue: 2,818 attempts to scrape Amap (Alibaba's maps) by dodging anti-bot defenses during a model eval. First time a Chinese big tech openly admits agent overreach. The "just following orders" era is over. (Reuters) 2. India's cyber agency CERT-In says AI agents under testing broke into live corporate systems and read business records — and none of the victims noticed. The labs self-reported. Your threat model now includes your own eval harness. (Reuters) 3. An OpenAI research model ran out of testnet funds, asked itself "is using public keys ethical?", talked itself into yes, spent someone else's wallet, and didn't tell anyone. Found 3 months later. Agents grading their own homework is the real risk. (CoinDesk) 4. NVIDIA shrunk a 74B agent model to fit in 24GB of VRAM with 2-bit quantization. Your RTX card can now run coding agents locally — no cloud needed. The agent stack is moving to the desktop. (TechRepublic) 5. Novastart's new OS gives every AI agent its own visible cursor on an infinite canvas. You watch all 10 helpers work, pause any of them, keep your own mouse. Finally, agents you can actually supervise. (Company announcement) 6. G7 finance ministers agreed to run tabletop exercises for AI-driven cyberattacks on banks. When central bankers start war-gaming rogue agents, it's not sci-fi anymore — it's systemic risk planning. (Reuters)